The multivariate methodology

B. Venkatesh

Anup Menon

WE CHOSE five factors for the multivariate models -- market capitalisation, past sales growth, past operating profit growth, past net profit growth and absolute earnings per share. The rationale for the choice is as follows:

Market capitalisation: Market capitalisation is an important indicator of the stock's liquidity. We did not include market capitalisation per se in the model. Instead, we used a dummy variable to model market capitalisation. Why? You would have observed that stock price was the dependent variable and market cap the independent one.

Click here for Table

Technically, market cap influences stock prices, and vice-versa. It, hence, gives rise to the statistical problem of simultaneity. We overcome this problem by using a dummy variable. The formulation of the variable is such that if the market cap of the stock is greater than the median market capitalisation of the entire sample, the variable is assigned a value of one; otherwise it is assigned a value of zero.

Sales growth: We feel that sales growth is an important variable, especially in cases where the industry is growing at a rapid pace. Why is this so? We all know that sales is the fundamental driver of cash inflows for any company.

It is, hence, an important factor for valuing stocks. Successful investment in the stock market, however, requires discerning the quality of the company in comparison to its peers. Taking absolute sales numbers may, therefore, be misleading. Hence, we used the growth rates in sales.

Operating profit growth: Operating profit growth reflects the company's ability to step up revenue without compromising on its pricing. That is, this variable helps investors discern whether the company is generating higher sales at a lower profit, or no profit, just to capture the market. We, therefore, took operating profit growth as the variable the market will look for in valuing tech stocks in conjunction with sales growth.

Net profit growth: Notwithstanding growth in sales and operating profits, a company's net profit growth may be negative. How? For instance, a company may have a high top-line growth but incur high research and development (R&D) expenditure, thereby, lowering net profits.

Under the circumstances, the market would prefer companies that can report high net profits even after accounting for R&D or other such items. We, however, chose to take net profit growth rate and not the absolute net profit value because it helps investors easily compare companies in the sector.

Methodology

For our analysis, we took the September and June 2000 and September 1999 quarterly earnings data for a sample of technology stocks. We used both simple and multivariate regression techniques. We preferred log-linear regression techniques to the standard ordinary least squares technique (OLS), because we believe that it is safe to assume that the data is not linear.

Using the log-linear method solves two critical modelling problems. On the one hand, it controls the statistical problem of `heteroskedacity' and, on the other, it helps normalise outlier observations.

To avoid problems associated with variables tending to zero, we added an arbitrary constant equivalent to one to all the variables. Further all variables outside the self-imposed two-standard deviation limit from the mean were excluded from the analysis. All estimates were calibrated at the 95 per cent confidence limit.

The dependent variable was so structured as to reduce any bias on account of `noise'. We estimated regressions using 50-day, 100-day and 200-day averages of the stock price.

For estimating the explanatory power of a traditional market model (one-factor model), we estimated the regression equation on the sample cross section with EPS as the independent variable and the 50-day stock price average as the dependent variable. We found that our model is statistically significant.

To test the explanatory power of the proposed five-factor model we ran a multiple regression on the cross-section sample. Similar to the earlier result, we found that the model was statistically significant. We also found that the intercept of the regression equation was statistically significant. This indicates that the five-factor model is superior in terms of explanatory to the traditional one-factor model.